HIM 400 Module 5 Data Mining Short Paper Example

Who Will Miss the Visit, and What Should We Do About It? Mining Appointment Data at Cold Brook Health

[Student Name]

Southern New Hampshire University

HIM 400: Communication and Technologies II

Module Five Short Paper

[Instructor Name]

[Date]

The organization, setting and figures below are a composite written as a model document. No real employer, client, colleague or patient is described.

Who Will Miss the Visit, and What Should We Do About It? Mining Appointment Data at Cold Brook Health

About one primary care appointment in eight at Cold Brook Health ends as a no-show. Each empty slot is a patient whose diabetes, blood pressure or depression went unchecked and a clinician's time that could have gone to someone on the waiting list. Operations leaders asked whether data mining could predict which appointments are likely to be missed. This paper describes how the analytics team built and tested such a model and why the most important decision was not which algorithm to use but what to do with a patient's risk score.

Data Mining as a Process

Data mining finds patterns in large data sets that can describe or predict behavior. Its main tasks include classification, which assigns records to categories such as attended or missed; clustering, which groups similar records without predefined labels; and association rules, which find items that occur together. Predicting no-shows is a classification task. The team followed the widely used CRISP-DM sequence: understand the business question, understand and prepare the data, build models, evaluate them and deploy the result. Writing down the business question first mattered, because it shaped every later choice, including which errors were acceptable.

Data and Predictors

The data set contained 148,300 scheduled primary care appointments from January 2024 through December 2025, of which 12.1% were no-shows. Pooling the published literature on missed appointments, Dantas et al. (2018) reported that rates varied widely across settings and that a patient's history of missed visits and the lead time between booking and the appointment were among the most frequently reported predictors. Guided by that review, the team built 14 candidate predictors from the scheduling system and the registry: prior no-shows in the past year, lead time in days, age, appointment type, day and hour, whether a reminder was confirmed, portal enrollment, insurance type and estimated travel distance. Clinical diagnoses were left out because the model's purpose was operational.

Building and Testing the Models

Rajkomar et al. (2019) caution that a model must be judged on data it has not seen, ideally from a later period, because patterns shift over time and a model tested on its own training data will look better than it is. The team therefore trained on 2024 appointments and tested on 2025. Two models were compared: logistic regression, which estimates how each predictor changes the odds of a no-show and is easy to explain, and a gradient-boosted tree model, which combines many small decision trees and can capture interactions. Discrimination on the 2025 test set, summarized as the area under the curve (AUC), reached 0.74 for logistic regression and 0.77 for the tree model. An AUC of 0.5 would mean the model ranks patients no better than a coin toss, and 1.0 would mean perfect separation. The small gain did not justify a model that staff could not interpret, so the team chose logistic regression.

What the Numbers Mean in Practice

A score becomes useful only when paired with a threshold and an action. The clinics can make about 15% of appointments the focus of extra attention each week, so the team flagged the highest-risk 15%. Table 1 shows what that means. Of flagged appointments, 31% were actually missed, compared with 12.1% overall, so outreach effort is concentrated where it is most needed. But the flags caught only 38% of all no-shows, so most missed appointments will still come as a surprise.

Table 1. Model Performance on 2025 Appointments at a 15% Flag Threshold

Note. Composite results from the logistic regression model.

Checking the Model by Subgroup

Obermeyer et al. (2019) examined a commercial risk tool used to enroll patients in care management and found that it assigned Black patients lower scores than white patients who were equally ill. The tool had been trained to forecast spending, and because unequal access had kept spending lower for Black patients, cost stood in poorly for need. The lesson is that a model can reproduce the conditions behind its training data. A no-show reflects barriers as much as choices, and at Cold Brook those barriers are concentrated among Medicaid patients and patients living more than 30 miles away. As expected, these groups were flagged far more often: 29% of appointments for patients on Medicaid and 34% for distant patients, against 15% overall. Accuracy was similar across groups, with areas under the curve between 0.72 and 0.75, so the model was not simply wrong for them. The question was what the flags would lead to.

Choosing the Action

Many clinics use no-show predictions to overbook. At Cold Brook, overbooking flagged slots would mean that patients on Medicaid and patients from distant towns, who already face the most barriers, would more often arrive to find a double-booked clinician and a long wait. Char et al. (2018) argue that the ethics of machine learning in health care depend on the intentions and incentives of those who design and deploy it, not only on the model's accuracy. The team recommended that flags trigger help instead: a reminder call from a scheduler who can offer a video visit, arrange a ride through the patient's insurance benefit or move the appointment to a nearer practice. Overbooking would not be used for flagged appointments.

Monitoring After Deployment

Once deployed, the model will be reviewed each quarter for three things: whether its accuracy holds on new appointments, whether flag rates by subgroup change and whether outreach actually lowers no-shows among flagged patients compared with the year before. Because the model predicts behavior that the outreach is designed to change, a successful program will make the model look less accurate over time, and the team will retrain it on new data each year.

Conclusion

Data mining gave Cold Brook a reasonably accurate way to identify appointments at risk of being missed. The more important work was deciding what a flag should mean. By testing the model on a later year, reporting results in plain terms, checking subgroups and linking predictions to help rather than overbooking, the team turned a technical result into a fair operational tool.

References

Char, D. S., Shah, N. H., & Magnus, D. (2018). Implementing machine learning in health care: Addressing ethical challenges. New England Journal of Medicine, 378(11), 981-983. https://doi.org/10.1056/NEJMp1714229

Dantas, L. F., Fleck, J. L., Cyrino Oliveira, F. L., & Hamacher, S. (2018). No-shows in appointment scheduling: A systematic literature review. Health Policy, 122(4), 412-421. https://doi.org/10.1016/j.healthpol.2018.02.002

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453. https://doi.org/10.1126/science.aax2342

Rajkomar, A., Dean, J., & Kohane, I. (2019). Machine learning in medicine. New England Journal of Medicine, 380(14), 1347-1358. https://doi.org/10.1056/NEJMra1814259